Create a response

OpenAI-compatible Responses API, passed through to the provider. Only models served directly by a provider that offers this endpoint can answer it — everything else returns 400 naming /v1/chat/completions instead. stream: true is supported. Of the built-in tools, web_search runs and is billed per search on top of tokens — the rate is on each model in GET /client/models (builtinTools), and each call shows its searches in the usage log. The other built-in tools (file_search, code_interpreter, hosted_shell, image_generation, computer_use) are refused. Function and custom tools work normally. Requires inference.